China

Existing video benchmarks often resemble image-based benchmarks, with questions like &quot;What color is the woman&#39;s dress in this video?&quot; or &quot;What&#39;s the weather like in the video?&quot; where models can answer using just a single frame or text, limiting the assessment of large vision language models (LVLMs) in terms of video perception and temporal reasoning abilities. To address this, we introduce GLIMPSE, a benchmark specifically designed to evaluate whether LVLMs truly understand video content. GLIMPSE includes 3,269 videos and over 4,342 highly visual-centric questions across 11 categories, such as Trajectory, Temporal QA, Forensics Detection, and Video Sorting, all meticulously curated and crafted by human annotators. Many of these questions require watching the entire video and reasoning based on video context to answer accurately, as they cannot be addressed using only a few frames or text information. Human testing on GLIMPSE achieves an average accuracy of 94.82%, but GLIMPSE poses significant challenges for existing LVLMs. Even the best-performing model, Gemini 1.5 Pro, reaches only 56.98% accuracy, indicating that current LVLMs still have a long way to go in truly understanding and interpreting video content.

EMNLP 2025

 GLIMPSE: Do Large Vision-Language Models Truly Think With Videos or Just Glimpse at Them?

video understanding

benchmark

## Welcome!
"I am excited to welcome you to this year’s edition of the Conference on Empirical Methods in Natural Language Processing! Importantly, it marks the 30th edition of EMNLP. With over 8,000 submissions, more than 3,000 accepted papers, and thousands of attendees, we have come a long way from that first
workshop, which had 14 accepted papers. As the field looks ahead, Suzhou is the fitting location for celebrating this milestone: rooted in a long literary tradition, yet modern and forward-looking, and home to a large share of the NLP community."<br>

*Message from the General Chair, Dirk Hovy*





<br>

Celebrate 30 Years of EMNLP! 
EMNLP 2025 will be held in Suzhou, China from November 5th to November 9th, 2025.

Existing video benchmarks often resemble image-based benchmarks, with questions like "What color is the woman's dress in this video?" or "What's the weather like in the video?" where models can answer using just a single frame or text, limiting the assessment of large vision language models (LVLMs) in terms of video perception and temporal reasoning abilities. To address this, we introduce GLIMPSE, a benchmark specifically designed to evaluate whether LVLMs truly understand video content. GLIMPSE includes 3,269 videos and over 4,342 highly visual-centric questions across 11 categories, such as Trajectory, Temporal QA, Forensics Detection, and Video Sorting, all meticulously curated and crafted by human annotators. Many of these questions require watching the entire video and reasoning based on video context to answer accurately, as they cannot be addressed using only a few frames or text information. Human testing on GLIMPSE achieves an average accuracy of 94.82%, but GLIMPSE poses significant challenges for existing LVLMs. Even the best-performing model, Gemini 1.5 Pro, reaches only 56.98% accuracy, indicating that current LVLMs still have a long way to go in truly understanding and interpreting video content.

technical paper

Reinforcement learning from human feedback (RLHF) and, at its core, reward modeling have become a crucial part of training powerful large language models (LLMs). One commonly overlooked factor in training high-quality reward models (RMs) is the effect of the base model, which is becoming more challenging to choose given the rapidly growing pool of LLMs. In this work, we present a systematic analysis of the effect of base model selection on reward modeling performance. Our results show that the performance can be improved by up to 14% compared to the most common (i.e., default) choice. Moreover, we showcase the strong statistical relation between some existing benchmarks and downstream performances. We also demonstrate that the results from a small set of benchmarks could be combined to boost the model selection (+18% on average in the top 5-10). Lastly, we illustrate the impact of different post-training steps on the final performance and explore using estimated data distributions to reduce performance prediction error.

A Systematic Analysis of Base Model Choice for Reward Modeling

Detecting self-contradictions within documents is a challenging task for ensuring textual coherence and reliability. While large language models (LLMs) have advanced in many natural language understanding tasks, document-level self-contradiction detection (DSCD) remains insufficiently studied. Recent approaches leveraging Chain-of-Thought (CoT) prompting aim to enhance reasoning and interpretability; however, they only gain marginal improvement and often introduce inconsistencies across repeated responses. We observe that such inconsistency arises from incomplete reasoning chains that fail to include all relevant contradictory sentences consistently. To address this, we propose a two-stage method that combines supervised fine-tuning (SFT) and reinforcement learning (RL) to enhance DSCD performance. In the SFT phase, a teacher model helps the model learn reasoning patterns, while RL further refines its reasoning ability. Our method incorporates a task-specific reward function to expand the model's reasoning scope, boosting both accuracy and consistency. On the ContraDoc benchmark, our approach significantly boosts Llama 3.1-8B-Instruct's accuracy from 38.5% to 51.1%, and consistency from 59.6% to76.2%.

Think Wider, Detect Sharper: Reinforced Reference Coverage for Document-Level Self-Contradiction Detection

Although prior work on bias mitigation has focused on promoting social equality and demographic parity, less attention has been given to aligning language model outputs with desired distributions. For example, we might want to align a model with real-world distributions to support factual grounding. Thus, we define bias as deviation from a desired distribution, which may be an equal or real-world distribution — depending on application goals. We propose a weighted adaptive loss based fine-tuning method that aligns LLM’s gender–profession output distribution with the desired distribution, while preserving language modeling performance. Using three profession sets—male-dominated, female-dominated, and gender-balanced— derived from U.S. labor statistics (2024), we evaluate both our adaptive method for reflecting reality and a non-adaptive variant for equality. Experiments on three masked language models and one autoregressive model (LLaMA 3.2-3B-Instruct) show near-complete bias mitigation under equality and about 30–75% (MLMs) and 50% (ALM) bias reduction, when aligning LLMs to distributions in the real world.

LLM Bias Detection and Mitigation through the Lens of Desired Distributions

Large language models (LLMs) have shown strong potential in complex reasoning tasks. However, as task complexity increases, their performance often degrades, resulting in hallucinations, errors, and logical inconsistencies. To enhance reasoning capabilities, Monte Carlo Tree Search (MCTS) has been introduced to guide the exploration of reasoning paths in a structured manner. Despite its advantages, traditional MCTS relies on fixed reasoning strategies, limiting the diversity of reasoning paths and the coverage of the solution space. To address these limitations, we propose Dynamic Strategy-Guided MCTS (DSG-MCTS), a novel framework that dynamically integrates multiple reasoning strategies, such as abductive and analogical reasoning, to expand the reasoning space. At the same time, DSG-MCTS enhances reasoning efficiency through a dynamic strategy selection mechanism that adapts to the task context. Experimental results on challenging reasoning benchmarks demonstrate that DSG-MCTS achieves improved accuracy and efficiency, outperforming existing state-of-the-art methods. The implementation is available at: https://anonymous.4open.science/r/DSG-MCTS-27AB.

DSG-MCTS: A Dynamic Strategy-Guided Monte Carlo Tree Search for Diversified Reasoning in Large Language Models

Can small language models (0.5B–5B parameters) meaningfully engage in trauma-informed, empathetic dialogue for individuals with PTSD? We answer this by introducing TIDE, a dataset of 10,000 two-turn dialogues across 500 diverse PTSD client personas, grounded in a three-factor empathy model: emotion recognition, distress normalization, and supportive reflection. All scenarios and reference responses were reviewed for realism and trauma sensitivity by a clinical psychologist specializing in PTSD. Eight small language models are evaluated before and after fine-tuning, with outputs compared to a frontier model (Claude Sonnet 3.5) as reference. Our IRB-approved human evaluation and automatic metrics reveal that, while fine-tuning generally improves perceived empathy, gains are highly scenario- and user-dependent, with smaller models facing an "empathy ceiling." Notably, demographic analyses show older adults value distress validation and graduate-educated users prefer nuanced replies, while gender effects are minimal. We highlight limitations of automatic metrics and the need for context- and user-aware system design. Our findings—along with the planned release of TIDE—offer a foundation for building safe, resource-efficient, and ethically sound empathetic AI to supplement, not replace, clinical mental health care.

The Pursuit of Empathy: Evaluating Small Language Models for PTSD Dialogue Support

Relational databases are central to modern data management, yet most data exists in unstructured forms like text documents. To bridge this gap, we leverage large language models (LLMs) to automatically synthesize a relational database by generating its schema and populating its tables from raw text. We introduce SQUiD, a novel neurosymbolic framework that decomposes this task into four stages, each with specialized techniques. Our experiments show that SQUiD consistently outperforms baselines across diverse datasets.

SQUiD: Synthesizing Relational Databases from Unstructured Text

Recently, Test-Time Scaling (TTS) has gained increasing attention for improving LLM reasoning performance at test time without retraining the model. A notable TTS technique is Self-Consistency (SC), which generates multiple reasoning chains in parallel and selects the final answer via majority voting. While effective, the order-of-magnitude computational overhead limits its broad deployment. Prior attempts to accelerate SC mainly rely on model-based confidence scores or heuristics with limited empirical support. For the first time, we theoretically and empirically analyze the inefficiencies of SC and reveal actionable opportunities for improvement. Building on these insights, we propose Slim-SC, a (thinking) step-wise pruning strategy that removes redundant chains using inter-chain similarity at the thought level. Experiments on three mathematical reasoning datasets and two recent LLM architectures confirm that Slim-SC not only reduces resource waste but also matches or even increases the accuracy of SC, providing a simple yet effective alternative for efficient test-time scaling with Self-Consistency.

Slim-SC: Thought Pruning for Efficient Scaling with Self-Consistency

Social text understanding is prone to fail when opinions are conveyed implicitly or sarcastically. It is therefore desired to model users' contexts in processing the texts authored by them. In this work, we represent users within a social embedding space that was learned from Twitter at large-scale. Similar to word embeddings that encode lexical semantics, the user embeddings encode general dimensions of social semantics. We conduct comprehensive experiments of author-informed stance prediction, showing improved learning generalization due to user modeling both within and across topics. Promising results are further shown for author-informed toxicity detection. The proposed approach may pave way to social NLP that considers user embeddings as contextual modality. Investigating the social factors that underlie author modeling, we find however that user stances as labeled in benchmark datasets exhibit socio-demographic patterns. Hence, author-informed stance detection approaches can both utilize, and should be aware of social biases.

 Towards Author-informed NLP: Mind the Social Bias

Large language models segment many words into multiple tokens, and companies that make those models claim that meaningful subword tokens are essential. To investigate whether subword tokens bear meaning, we segmented tens of thousands of words from each of 41 languages according to three generations of GPT tokenizers. We found that words which share tokens are more semantically similar than expected by chance or expected from length alone, that tokens capture morphological information even when they don’t look like morphemes, and that tokens capture more information than is explained by morphology. In languages that use a script other than the Latin alphabet, GPT-4 tokens are uninformative, but GPT-4o has improved this situation. These results suggest that comparing tokens to morphemes overlooks the wider variety of semantic information available in word form and that standard tokenization methods successfully capture much of that information.

How much Semantic Information is Available in Large Language Model Tokens

Pre-trained language models (PLMs) have driven substantial progress in natural language processing but remain vulnerable to adversarial attacks, raising concerns about their robustness in real-world applications. Previous studies have sought to mitigate the impact of adversarial attacks by introducing adversarial perturbations into the training process, either implicitly or explicitly. While both strategies enhance robustness, they often incur high computational costs. In this work, we propose a simple yet effective add-on module that enhances the adversarial robustness of PLMs by removing instance-level principal components, without relying on conventional adversarial defences or perturbing the original training data. Our approach transforms the embedding space to approximate Gaussian properties, thereby reducing its susceptibility to adversarial perturbations while preserving semantic relationships. This transformation aligns embedding distributions in a way that minimises the impact of adversarial noise on decision boundaries, enhancing robustness without requiring adversarial examples or costly training-time augmentation. Evaluations on eight benchmark datasets show that our approach improves adversarial robustness while maintaining comparable before-attack accuracy to baselines, achieving a balanced trade-off between robustness and generalisation.

Downloads

Next from EMNLP 2025

A Systematic Analysis of Base Model Choice for Reward Modeling

Similar lecture

Video-RTS: Rethinking Reinforcement Learning and Test-Time Scaling for Efficient and Enhanced Video Reasoning